Deep learning based denoising of Lu-177 SPECT/CT imaging

Ziel/Aim CNN-based denoising is increasingly applied to suppress noise in PET images [1]. This work evaluates these methods for application in SPECT imaging based on a previously presented large dataset of 10,000 Monte Carlo (MC) simulated SPECT acquisitions [2].

Methodik/Methods The generation of the dataset consisting of 10,000 random activity distributions and the associated SPECT simulations (“noise-free data”) are described in [2]. Anthropomorphic XCAT phantoms were used as attenuating and scattering medium to ensure patient-realistic geometries. Finally, Poisson noise was added (“noisy data”). Reconstructions were performed using CASToR-OSEM with attenuation and scatter correction (2 subsets, 10 iterations). Denoising in projection space and after reconstruction was performed using three CNN-based techniques (9000/500/500 for training/validation/testing) [1]: Noise2Clean (noisy data → noise-free data); Noise2Noise (noisy data → noisy data with a different noise realization); Noisier2Noise (noisy data with additional Poisson noise → noisy data). In addition, post-reconstruction denoising with a Gaussian kernel (FWHM: 8mm) was performed. The different methods were compared using the SSIM, NRMSE and PSNR similarity measures.

Ergebnisse/Results The best results were achieved by Noise2Clean in the projection space (SSIM: 0.998, NRMSE: 0.94%, PSNR: 40.98). All CNN-based denoising outperformed conventional Gaussian denoising (0.992, 1.90%, 34.63). Additionally, denoising projections significantly outperformed denoising of reconstructed images (paired signed Wilcoxon-Test, p Noise2Noise>Noisier2Noise.

Schlussfolgerungen/Conclusions This study shows that CNNs are promising for denoising SPECT images. All investigated CNN-based methods showed better performance than Gaussian denoising. Additionally, denoising in projection space provided better results than after SPECT reconstruction.

Publikationsverlauf

Artikel online veröffentlicht:
30. März 2023

• Literatur/References

• 1 Yie et al. Nucl Med Mol Imaging. 54. 2020
• 2 Leube et al. EJNMMI Phys. 9. 2022